For example, an X-ray chest image has a very broad range of pixel values since it is made up of an image region of lungs through which X-rays are readily transmitted, and an image region of a mediastinal part through which X-rays are hardly transmitted. For this reason, it has been considered to be difficult to obtain an X-ray chest image that allows to simultaneously observe both the lungs and mediastinal part.
As a method of avoiding this problem, a method described in SPIE Vol. 626 MedicineXIV/PACSIV (1986) is known. This method is described using constants A, B, and C (for example, A=3, B=0.7) by:SD=A[Sorg−SUS]+B[SUS]+C  (1)where SD is the pixel value of an image after processing, Sorg is the pixel value (input pixel value) of an original image (input image), and SUS is the pixel value of a low-frequency image of the original image.
This method can change weighting coefficients for high-frequency components (first term) and low-frequency components (second term). For example, when A=3 and B=0.7, the effect of emphasizing the high-frequency components and compressing the overall dynamic range can be obtained. Five radiologists evaluated that this method is effective for diagnosis compared to an image without any processing.
If the ratio of A is increased in equation (1), the ratio of high-frequency components increases, and a sharpening effect can be obtained. On the other hand, if the ratio of B is changed, the magnitudes of low-frequency components are changed as well as the dynamic range of the image SD.
Japanese Patent No. 2509503 describes a method which is described by:SD=Sorg+F[G(Px, Py)]  (2)where SD is the pixel value after processing, Sorg is the original pixel value (input pixel value), Py is the average profile of a plurality of Y-profiles of an original image, and Px is the average profile of a plurality of X-profiles.
The characteristics of the function F(x) will be explained below. If “x>Dth”, F(x) becomes “0”. If “0≦x≦Dth”, F(x) monotonously decreases to have “E” as a segment and “E/Dth” as a slope. F(x) is given by:F(x)=E−(E/Dth)x, when 0≦x≦Dth =0, when x>Dth  (3)Py=(ΣPyi)/n  (4)Px=(ΣPxi)/n  (5)where (i=1 to n), and Pyi and Pxi are profiles. For example, G(Px, Py) is given by:G(Px, Py)=max(Px, Py)  (6)In this method, of the pixel value (density value) range of the original image, the pixel value (density value) range in which the pixel values of a low-frequency image are equal to or smaller than Dth is compressed.
As a method similar to the method of Japanese Patent No. 2509503, a method described in “Anan et. al., Japanese Journal of Radiological Technology, Vol. 45, No. 8, August 1989, p. 1030”, and Japanese Patent No. 2663189 is known. Using the monotone decreasing function f(x), this method is described by:SD=Sorg+f(SUS)  (7)SUS=ΣSorg/M2  (8)where SD is the pixel value after processing, Sorg is the original pixel value, and SUS is the average pixel value upon calculating a moving average using a mask size M×M pixels in the original image.
In this method, the low-frequency image generation method is different from that in the method given by equation (2). In the method given by equation (2), a low-frequency image is generated based on one-dimensional data, while in this method, a low-frequency image is generated based on two-dimensional data. In this method as well, of the pixel value (density value) range of the original image, the pixel value (density value) range in which the pixel values of a low-frequency image are equal to or smaller than Dth is compressed.
The aforementioned dynamic range compression method can be expressed using a function f1( ) of converting a low-frequency image by:SD=f1(SUS)+(Sorg−SUS)  (9)Note that the variable of a function may be omitted for the sake of simplicity in this specification.
In equation (9), the dynamic range is changed by changing low-frequency components using the function f1( ). The dynamic range compression method given by equation (9) will be explained below. FIGS. 1 and 2 are views for explaining the principle of that method. The uppermost view in FIG. 1 shows the profile of an edge portion of an original image, the middle view shows the profile of a smoothed image of that original image, and the lowermost view shows the profile of a high-frequency image generated by subtracting the smoothed image from the original image. In FIG. 2, the uppermost view shows the profile of an image obtained by multiplying by ½ the absolute values of the smoothed image in the middle view of FIG. 1, the middle view shows the same profile as that of the high-frequency image in FIG. 1, and the lowermost view shows the profile of an image obtained by adding the high-frequency image in the interrupt view to the image in the uppermost view obtained by converting the values of the smoothed image. A process for obtaining an image, the dynamic range of which is compressed, like the image shown in the lowermost view in FIG. 2, is called a dynamic range compression process.
As can be seen from FIG. 1, the smoothed image cannot maintain an edge structure in the edge portion, and the high-frequency components have large values at the edge portion. Note that the source original image can be recovered by adding the smoothed image and high-frequency image.
However, as shown in FIG. 2, when the high-frequency image is added to the image obtained by converting the values of the low-frequency image, the edge structure collapses, as indicated by arrows in FIG. 2. Such phenomenon is called overshoot/undershoot (to be also referred to as overshoot, overshoot, or the like hereinafter).
Note that equation (10) changes the original image by the function f1( ), and expresses normal tone conversion; it can change the dynamic range of the overall original image.SD=f1(Sorg)  (10)
In recent years, a multiple-frequency process (to be also referred to as a multiple-frequency transformation process hereinafter) using Laplacian pyramid transformation and wavelet transformation has been developed. In such multiple-frequency process, high-frequency components such as Laplacian coefficients or wavelet coefficients (to be referred to as frequency coefficients hereinafter) obtained by decomposing an image into frequency components are converted using a nonlinear function shown in FIG. 3 or 4. In FIGS. 3 and 4, the abscissa plots the input coefficients, and the ordinate plots the output coefficients. FIGS. 3 and 4 show conversion curves when the coefficients are +, and the same conversion is made even when the coefficients are −. That is, FIGS. 3 and 4 show only the first quadrant of an odd function. In this specification, all functions used to convert frequency coefficients are odd functions, and only their first quadrants are shown. Also, “curve” and “functions” may be used as equivalent terms. FIG. 3 shows a monotone increasing concave function (upward convex). When coefficients are converted using such function form, coefficients can be increased in a small coefficient region, and the coefficients can be saturated in a large coefficient region. Therefore, when the small coefficient region expresses effective image components of, e.g., a fine structure, an image process that emphasizes the fine structure is done. In addition, since the coefficients of the large coefficient region are saturated, an effect of suppressing emphasis of an edge structure or the like can be obtained.
The curve form shown in FIG. 4 is used in a method called degeneracy of wavelet, and converts frequency coefficients less than a predetermined absolute value (threshold value) 3001 shown in FIG. 4 into 0 (zero), thus providing an effect of suppressing noise.
Furthermore, a method of changing the dynamic range of a recovered image by changing coefficients in the lowermost-frequency band in the multiple-frequency process is known.
Along with the advance of digital technologies in recent years, a radiation image such as an X-ray image or the like is converted into a digital signal, such digital image undergoes an image process, and the processed image is displayed on a display device (e.g., a CRT, liquid crystal display, or the like) or is recorded on a recording medium such as a film or the like by a recording apparatus (printer or the like). Such image process is categorized into a pre-process for correcting an image obtained from an image sensing device depending on the characteristics or the like of the image sensing device, and a quality assurance (QA) process for converting the image (original image) that has undergone the pre-process into an image with image quality suitable for diagnosis. Of these processes, the QA process includes frequency processes such as a sharpening process for emphasizing the high-frequency components of an original image, a noise reduction process for suppressing high-frequency components, and the like.
The sharpening process is based on the following process. That is, a high-frequency image as high-frequency components of an original image shown in FIG. 5C is generated by subtracting a blurred image (smoothed image) as low-frequency components of the original image shown in FIG. 5B from the original image (including an edge portion) shown in FIG. 5A. Then, as shown in FIG. 6, the high-frequency image is added to the original image to obtain an image with higher sharpness (sharpened image). FIGS. 5A, 5B, and 5C are waveform charts for explaining the sharpening process, in which FIG. 5A is a waveform chart showing the profile of the original image including an edge portion, FIG. 5B is a waveform chart showing the profile of the smoothed image obtained by smoothing the original image shown in FIG. 5A, and FIG. 5C is a waveform chart showing the profile of the high-frequency image generated by subtracting the smoothed image shown in FIG. 5B from the original image shown in FIG. 5A. FIG. 6 is a waveform chart showing the profile of the sharpened image obtained by adding the high-frequency image shown in FIG. 5C to the original image shown in FIG. 5A.
With the dynamic range compression process given by equation (1), since high- and low-frequency components are converted by uniformly multiplying by different constants, the dynamic range compression process can be achieved, but overshoot occurs.
The dynamic range compression process given by equation (2) has no disclosure about an idea that adjusts high-frequency components, and changes only low-frequency components. Hence, the dynamic range compression process can be achieved, but overshoot occurs.
With the dynamic range compression process which is given by equation (9) and adds high-frequency components to the converted smoothed image (low-frequency components), only low-frequency components are converted, and high-frequency components remain the same. Hence, overshoot occurs again.
For example, when the entire smoothed image is converted to be ½ in FIG. 2, if high-frequency components of portions corresponding to overshoot and undershoot are multiplied by ½, the edge structure is preserved in the image of the dynamic range compression process. However, when the entire smoothed image is converted to be ⅓ or is converted using a complicated curve form, overshoot or undershoot occurs if the high-frequency components of portions corresponding to overshoot and undershoot are multiplied by ½.
As a method of suppressing such overshoot and undershoot, the present applicant has filed Japanese Patent Laid-Open No. 2000-316090. This method suppresses overshoot and undershoot by suppressing high-frequency component values corresponding to overshoot and undershoot portions. However, such method of suppressing portions with large high-frequency value portions can suppress overshoot and undershoot, but cannot perfectly preserve the edge structure. Therefore, portions where high-frequency components are suppressed become unnatural.
On the other hand, the edge structure can be perfectly preserved if high- and low-frequency components are changed at the same ratio, like in FIG. 1 in which the original image can be recovered by adding the high-frequency image and smoothed image. However, such method is nothing but tone conversion given by equation (10). Since simple tone conversion can adjust the dynamic range but cannot adjust frequency components, when, for example, the dynamic range is compressed, a fine structure or the like undesirably flattens out. Also, the effect of the sharpening process or the like cannot be obtained.
Upon converting the frequency coefficients in the multiple-frequency process using the conversion curve shown in FIG. 3, overshoot can be suppressed by the same effect as in Japanese Patent Laid-Open No. 2000-316090. However, the edge structure cannot be perfectly preserved, either, as described above, and an unnatural edge portion may appear.
Also, when the coefficients of the lowest-frequency band are changed, the edge structure cannot be preserved due to the same principle as described above, and overshoot may be generated. That is, when the absolute values of coefficients of a partial frequency band which forms the edge portion are changed, the structure of the edge portion collapses somehow, thus generating unnaturalness (artifacts).
If all the frequency coefficients are changed at an identical ratio, the edge structure can be prevented from collapsing. However, such change is merely tone conversion. Hence, no effect of the frequency process is expected.
When coefficients are converted using the conversion curve shown in FIG. 4, the edge structure is preserved in an inversely converted image (e.g., an image that has undergone inverse wavelet transformation). However, since there is no idea of emphasizing coefficients, the sharpening effect cannot be obtained by the inversely converted image at all. When the slope of the curve in FIG. 4 is set other than 1, the edge structure is not preserved, and overshoot or the like occurs.
On the other hand, the conventional sharpening process cannot sharpen an image that contains an edge portion with high quality. For example, as best illustrated in FIG. 7, since high-frequency components of the edge portion have values extremely larger than those of other portions, extremely protruding regions (regions a and b indicated by circles in FIG. 7) may appear in a sharpened image obtained by adding them. Note that FIG. 7 is a waveform chart showing the profile of a sharpened image that suffers overshoot. These regions a and b are artifacts called overshoot (region b may also be called undershoot). Such image, the edge portion of which is excessively emphasized by overshoot, is unnatural. Especially, in case of a medical image such as a radiation image or the like which is used in diagnosis, generation of such artifacts is not preferable. On the other hand, in order to suppress overshoot, a high-frequency image may be decreased at a predetermined ratio, and may then be added to an original image. However, the effect of the sharpening process undesirably lowers in a region other than the edge portion.
In the above description, when high-frequency components are emphasized while preserving low-frequency components of an image, the edge structure collapses (the same applies to suppression of low-frequency components while preserving high-frequency components). Conversely, when high-frequency components are suppressed while preserving low-frequency components of an image, the edge structure collapses again (the same applies to emphasis of low-frequency components while preserving high-frequency components). In this case, in place of overshoot, sharpness of the edge portion is lost, and the edge structure collapses with the edge portion being blurred.